Presenting the new data set

I made all the changes that I need. Now my data is ready. let’s take a look at it.

head(HR.new)

##           ï..Employee_Name EmpID MarriedID MaritalStatusID GenderID EmpStatusID
## 1      Adinolfi, Wilson  K 10026         0               0        1           1
## 2 Ait Sidi, Karthikeyan    10084         1               1        1           5
## 3        Akinkuolie, Sarah 10196         1               1        0           5
## 4             Alagbe,Trina 10088         1               1        0           1
## 5         Anderson, Carol  10069         0               2        0           5
## 6        Anderson, Linda   10002         0               0        0           1
##   DeptID PerfScoreID FromDiversityJobFairID        Salary Termd PositionID
## 1      5           4                      0       average     0         19
## 2      3           3                      0          high     1         27
## 3      5           3                      0          good     1         20
## 4      5           3                      0          good     0         19
## 5      5           3                      0 below average     1         19
## 6      5           4                      0       average     0         19
##    Position State  Zip      DOB Sex MaritalDesc CitizenDesc HispanicLatino
## 1 Personnel    MA 1960 07/10/83  M       Single  US Citizen             No
## 2 Personnel    MA 2148 05/05/75  M      Married  US Citizen             No
## 3 Personnel    MA 1810 09/19/88   F     Married  US Citizen             No
## 4 Personnel    MA 1886 09/27/88   F     Married  US Citizen             No
## 5 Personnel    MA 2169 09/08/89   F    Divorced  US Citizen             No
## 6 Personnel    MA 1844 05/22/77   F      Single  US Citizen             No
##   RaceDesc DateofHire DateofTermination        TermReason
## 1    White   7/5/2011                   N/A-StillEmployed
## 2    White  3/30/2015         6/16/2016     career change
## 3    White   7/5/2011         9/24/2012             hours
## 4    White   1/7/2008                   N/A-StillEmployed
## 5    White  7/11/2011          9/6/2016  return to school
## 6    White   1/9/2012                   N/A-StillEmployed
##         EmploymentStatus        Department    ManagerName ManagerID
## 1                 Active Production        Michael Albert        22
## 2 Voluntarily Terminated             IT/IS     Simon Roup         4
## 3 Voluntarily Terminated Production        Kissy Sullivan        20
## 4                 Active Production          Elijiah Gray        16
## 5 Voluntarily Terminated Production        Webster Butler        39
## 6                 Active Production              Amy Dunn        11
##   RecruitmentSource PerformanceScore EngagementSurvey EmpSatisfaction
## 1          LinkedIn          Exceeds     most engaged       satisfied
## 2            Indeed      Fully Meets     most engaged    dissatisfied
## 3          LinkedIn      Fully Meets    least engaged    dissatisfied
## 4            Indeed      Fully Meets     most engaged       satisfied
## 5     Google Search      Fully Meets     most engaged       satisfied
## 6          LinkedIn          Exceeds     most engaged       satisfied
##   SpecialProjectsCount LastPerformanceReview_Date DaysLateLast30  Absences
## 1                    0                  1/17/2019              0   perfect
## 2                    6                  2/24/2016              0 concerned
## 3                    0                  5/15/2012              0   perfect
## 4                    0                   1/3/2019              0 concerned
## 5                    0                   2/1/2016              0   perfect
## 6                    0                   1/7/2019              0 concerned

Training a model on the data

Below we demonstrate association rule mining with apriori() function from arules package which we already called. Apriori counts transactions to find frequent itemsets and then derive association rules from them. With settings of: supp=0.05, which is the minimum support of rules; conf=0.25, which is the minimum confidence of rules; and minlen=2 & maxlen=5, which are the minimum and the maximum length of rules, here is our rule:

myrule <- apriori(datatrans, parameter = list(minlen=2, maxlen=5,support=0.05,  confidence=0.25))

## Apriori
## 
## Parameter specification:
##  confidence minval smax arem  aval originalSupport maxtime support minlen
##        0.25    0.1    1 none FALSE            TRUE       5    0.05      2
##  maxlen target  ext
##       5  rules TRUE
## 
## Algorithmic control:
##  filter tree heap memopt load sort verbose
##     0.1 TRUE TRUE  FALSE TRUE    2    TRUE
## 
## Absolute minimum support count: 15 
## 
## set item appearances ...[0 item(s)] done [0.00s].
## set transactions ...[25 item(s), 303 transaction(s)] done [0.00s].
## sorting and recoding items ... [20 item(s)] done [0.00s].
## creating transaction tree ... done [0.00s].
## checking subsets of size 1 2 3 4 5

## Warning in apriori(datatrans, parameter = list(minlen = 2, maxlen = 5, support
## = 0.05, : Mining stopped (maxlen reached). Only patterns up to a length of 5
## returned!

##  done [0.00s].
## writing ... [1632 rule(s)] done [0.00s].
## creating S4 object  ... done [0.00s].

myrule

## set of 1632 rules

myrule object contains a set of 1632 association rules.

Evaluating model performance

summary(myrule)

## set of 1632 rules
## 
## rule length distribution (lhs + rhs):sizes
##   2   3   4   5 
## 192 619 622 199 
## 
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   2.000   3.000   4.000   3.507   4.000   5.000 
## 
## summary of quality measures:
##     support          confidence        coverage            lift       
##  Min.   :0.05281   Min.   :0.2500   Min.   :0.05281   Min.   :0.6613  
##  1st Qu.:0.05941   1st Qu.:0.3556   1st Qu.:0.09571   1st Qu.:0.9914  
##  Median :0.07591   Median :0.6137   Median :0.16832   Median :1.0834  
##  Mean   :0.10156   Mean   :0.5870   Mean   :0.20117   Mean   :1.1612  
##  3rd Qu.:0.11634   3rd Qu.:0.8000   3rd Qu.:0.24092   3rd Qu.:1.2000  
##  Max.   :0.64686   Max.   :1.0000   Max.   :0.81518   Max.   :7.8194  
##      count       
##  Min.   : 16.00  
##  1st Qu.: 18.00  
##  Median : 23.00  
##  Mean   : 30.77  
##  3rd Qu.: 35.25  
##  Max.   :196.00  
## 
## mining info:
##       data ntransactions support confidence
##  datatrans           303    0.05       0.25

In the part of ‘summary of quality measures’ we can see that the max support is 0.646 which represent the most frequent item. In our case it’s Position=Personnel. The max count is 196 which also present the number of employees who are in Personnel position. The lift is a measure of how much more likely one item is to be present(exist) relative to its typical present(existing) rate. In this rule the maximum lift is 7.

Determine the factors that influence the variable EmpSatisfaction

We are interested in only rules with rhs indicating employee satisfaction , so I set ’ rhs=c(“EmpSatisfaction=dissatisfied”, “EmpSatisfaction=satisfied”) ’ in appearance to make sure that only “EmpSatisfaction=dissatisfied” and “EmpSatisfaction=satisfied” will appear in the right hand side of rules. All other items can appear in the left hand side, as set with default=“lhs”. Rules are sorted by lift to make high-lift rules appear first.

rule <- apriori(datatrans, parameter = list(minlen=2, maxlen=5,support=0.05,  confidence=0.25),appearance= list(rhs=c("EmpSatisfaction=dissatisfied", "EmpSatisfaction=satisfied"), default="lhs"), control = list(verbose=F))
rule

## set of 245 rules

inspect(sort(rule, by = "lift")[1:9]) #Sorting the set of association rules

##     lhs                                 rhs                               support confidence   coverage     lift count
## [1] {Salary=average,                                                                                                  
##      EngagementSurvey=least engaged} => {EmpSatisfaction=dissatisfied} 0.05940594  0.6206897 0.09570957 1.621284    18
## [2] {EngagementSurvey=least engaged,                                                                                  
##      Absences=perfect}               => {EmpSatisfaction=dissatisfied} 0.06600660  0.5882353 0.11221122 1.536511    20
## [3] {Position=Personnel,                                                                                              
##      EngagementSurvey=least engaged,                                                                                  
##      Absences=perfect}               => {EmpSatisfaction=dissatisfied} 0.05280528  0.5333333 0.09900990 1.393103    16
## [4] {Salary=below average,                                                                                            
##      Department=Production       ,                                                                                    
##      PerformanceScore=Fully Meets,                                                                                    
##      Absences=good}                  => {EmpSatisfaction=satisfied}    0.05280528  0.8000000 0.06600660 1.296257    16
## [5] {Salary=average,                                                                                                  
##      EngagementSurvey=most engaged}  => {EmpSatisfaction=satisfied}    0.06270627  0.7916667 0.07920792 1.282754    19
## [6] {Salary=average,                                                                                                  
##      Position=Personnel,                                                                                              
##      EngagementSurvey=most engaged}  => {EmpSatisfaction=satisfied}    0.06270627  0.7916667 0.07920792 1.282754    19
## [7] {Salary=average,                                                                                                  
##      Department=Production       ,                                                                                    
##      EngagementSurvey=most engaged}  => {EmpSatisfaction=satisfied}    0.05940594  0.7826087 0.07590759 1.268077    18
## [8] {Salary=average,                                                                                                  
##      Position=Personnel,                                                                                              
##      Department=Production       ,                                                                                    
##      EngagementSurvey=most engaged}  => {EmpSatisfaction=satisfied}    0.05940594  0.7826087 0.07590759 1.268077    18
## [9] {PerformanceScore=Fully Meets,                                                                                    
##      EngagementSurvey=engaged,                                                                                        
##      Absences=good}                  => {EmpSatisfaction=satisfied}    0.08250825  0.7812500 0.10561056 1.265876    25

245 rules satisfy those criteria.

From the first rule we can be 62% sure that when an employee is least engaged and have an average salary is dissatisfaction. This rule is involved in 5.94% of the entire transactions, and the lift implies that an employees who have these characteristics are 1.62 times more likely to be dissatisfied than the typical employee.

In the above result, the 7th and 8th rules show that we can be 78,2% sure that an employees who are most engaged, have and average salary and work in production department are of the same satisfaction degree, with lift of 1,268 which indicates employees with same characteristics are 1,268 likely to be satisfied.

Taking subsets of association rules

SatisfiedRule=subset(rule, items  %in% "EmpSatisfaction=satisfied")
inspect(sort(SatisfiedRule, by = "lift")[7:11])

##     lhs                               rhs                            support confidence   coverage     lift count
## [1] {Salary=average,                                                                                             
##      Absences=good}                => {EmpSatisfaction=satisfied} 0.05610561  0.7727273 0.07260726 1.252066    17
## [2] {Salary=average,                                                                                             
##      EngagementSurvey=engaged}     => {EmpSatisfaction=satisfied} 0.05610561  0.7727273 0.07260726 1.252066    17
## [3] {Position=Personnel,                                                                                         
##      PerformanceScore=Fully Meets,                                                                               
##      EngagementSurvey=engaged,                                                                                   
##      Absences=good}                => {EmpSatisfaction=satisfied} 0.06600660  0.7692308 0.08580858 1.246401    20
## [4] {Salary=below average,                                                                                       
##      PerformanceScore=Fully Meets,                                                                               
##      Absences=good}                => {EmpSatisfaction=satisfied} 0.05280528  0.7619048 0.06930693 1.234530    16
## [5] {Salary=below average,                                                                                       
##      Position=Personnel,                                                                                         
##      PerformanceScore=Fully Meets,                                                                               
##      Absences=good}                => {EmpSatisfaction=satisfied} 0.05280528  0.7619048 0.06930693 1.234530    16

DisatRule=subset(rule, items %in% "EmpSatisfaction=dissatisfied")  
DisatRule

## set of 87 rules

inspect(sort(DisatRule, by = "lift")[9:13])

##     lhs                                 rhs                               support confidence  coverage     lift count
## [1] {EngagementSurvey=least engaged,                                                                                 
##      Absences=concerned}             => {EmpSatisfaction=dissatisfied} 0.06600660  0.4444444 0.1485149 1.160920    20
## [2] {Position=Personnel,                                                                                             
##      PerformanceScore=Fully Meets,                                                                                   
##      EngagementSurvey=least engaged} => {EmpSatisfaction=dissatisfied} 0.08250825  0.4385965 0.1881188 1.145644    25
## [3] {Position=Personnel,                                                                                             
##      Department=IT/IS,                                                                                               
##      PerformanceScore=Fully Meets}   => {EmpSatisfaction=dissatisfied} 0.05610561  0.4358974 0.1287129 1.138594    17
## [4] {Salary=good,                                                                                                    
##      Department=Production       }   => {EmpSatisfaction=dissatisfied} 0.07590759  0.4339623 0.1749175 1.133539    23
## [5] {Salary=high,                                                                                                    
##      Position=Personnel,                                                                                             
##      PerformanceScore=Fully Meets}   => {EmpSatisfaction=dissatisfied} 0.06270627  0.4318182 0.1452145 1.127939    19

Rules visualization for:

plot(DisatRule, engine = "htmlwidget")

## To reduce overplotting, jitter is added! Use jitter = 0 to prevent jitter.

plot(DisatRule[50:80] , method = "paracoord")

plot(SatisfiedRule, method = "graph", engine = "htmlwidget")

## Warning: Too many rules supplied. Only plotting the best 100 rules using lift
## (change control parameter max if needed)

Computing the critical value at 5% alpha level

alpha = .05
z1 = qnorm(1-alpha)
z1 

## [1] 1.644854

ifelse(z > z1,"Reject the Null Hypothesis","Accept the Null Hypothesis")

## [1] "Reject the Null Hypothesis"

This proves what we’ve already noticed earlier on the graph.

Define full and null models and do step procedure

model.null = glm(Termd ~ 1,family = "binomial", train)

model.full = glm(Termd ~  MaritalStatusID  + DeptID  + PerfScoreID   + Salary   + MaritalDesc   +  EngagementSurvey  + EmpSatisfaction +  Absences + SpecialProjectsCount  +  DaysLateLast30  +  PerformanceScore + CitizenDesc +  HispanicLatino  + RaceDesc, family ="binomial", train)

step(model.null,
     scope = list(upper=model.full),
     direction="both",
     data=train)

## Start:  AIC=290.04
## Termd ~ 1
## 
##                        Df Deviance    AIC
## + DaysLateLast30        1   283.40 287.40
## + SpecialProjectsCount  1   283.46 287.46
## + Salary                1   284.66 288.66
## + DeptID                1   284.97 288.97
## + MaritalDesc           4   279.40 289.40
## + PerfScoreID           1   285.81 289.81
## <none>                      288.04 290.04
## + Absences              1   286.30 290.30
## + MaritalStatusID       1   287.16 291.16
## + CitizenDesc           2   285.74 291.74
## + PerformanceScore      3   284.02 292.02
## + EmpSatisfaction       1   288.03 292.03
## + EngagementSurvey      1   288.04 292.04
## + HispanicLatino        2   287.15 293.15
## + RaceDesc              4   283.78 293.78
## 
## Step:  AIC=287.4
## Termd ~ DaysLateLast30
## 
##                        Df Deviance    AIC
## + SpecialProjectsCount  1   279.49 285.49
## + EngagementSurvey      1   279.68 285.68
## + MaritalDesc           4   274.01 286.01
## + Salary                1   280.65 286.65
## + DeptID                1   280.88 286.88
## <none>                      283.40 287.40
## + Absences              1   281.69 287.69
## + PerformanceScore      3   277.86 287.86
## + MaritalStatusID       1   282.06 288.06
## + CitizenDesc           2   280.96 288.96
## + EmpSatisfaction       1   283.15 289.15
## + PerfScoreID           1   283.32 289.32
## - DaysLateLast30        1   288.04 290.04
## + HispanicLatino        2   282.63 290.63
## + RaceDesc              4   279.56 291.56
## 
## Step:  AIC=285.49
## Termd ~ DaysLateLast30 + SpecialProjectsCount
## 
##                        Df Deviance    AIC
## + MaritalDesc           4   270.06 284.06
## + EngagementSurvey      1   276.32 284.32
## <none>                      279.49 285.49
## + Absences              1   277.83 285.83
## + PerformanceScore      3   273.96 285.96
## + MaritalStatusID       1   278.42 286.42
## + CitizenDesc           2   276.47 286.47
## + Salary                1   279.06 287.06
## + EmpSatisfaction       1   279.21 287.21
## - SpecialProjectsCount  1   283.40 287.40
## + PerfScoreID           1   279.41 287.41
## - DaysLateLast30        1   283.46 287.46
## + DeptID                1   279.49 287.49
## + HispanicLatino        2   278.82 288.82
## + RaceDesc              4   275.39 289.39
## 
## Step:  AIC=284.06
## Termd ~ DaysLateLast30 + SpecialProjectsCount + MaritalDesc
## 
##                        Df Deviance    AIC
## + EngagementSurvey      1   266.93 282.93
## <none>                      270.06 284.06
## + PerformanceScore      3   264.08 284.08
## + CitizenDesc           2   266.64 284.64
## + Absences              1   268.78 284.78
## - MaritalDesc           4   279.49 285.49
## + Salary                1   269.78 285.78
## + EmpSatisfaction       1   269.84 285.84
## - SpecialProjectsCount  1   274.01 286.01
## + DeptID                1   270.01 286.01
## + PerfScoreID           1   270.03 286.03
## - DaysLateLast30        1   274.66 286.66
## + HispanicLatino        2   269.31 287.31
## + RaceDesc              4   266.34 288.34
## 
## Step:  AIC=282.93
## Termd ~ DaysLateLast30 + SpecialProjectsCount + MaritalDesc + 
##     EngagementSurvey
## 
##                        Df Deviance    AIC
## <none>                      266.93 282.93
## + Absences              1   265.57 283.57
## + CitizenDesc           2   263.62 283.62
## + PerformanceScore      3   261.87 283.87
## - EngagementSurvey      1   270.06 284.06
## - MaritalDesc           4   276.32 284.32
## - SpecialProjectsCount  1   270.33 284.33
## + Salary                1   266.56 284.56
## + DeptID                1   266.63 284.63
## + EmpSatisfaction       1   266.77 284.77
## + PerfScoreID           1   266.82 284.82
## + HispanicLatino        2   265.99 285.99
## + RaceDesc              4   262.96 286.96
## - DaysLateLast30        1   274.66 288.66

## 
## Call:  glm(formula = Termd ~ DaysLateLast30 + SpecialProjectsCount + 
##     MaritalDesc + EngagementSurvey, family = "binomial", data = train)
## 
## Coefficients:
##          (Intercept)        DaysLateLast30  SpecialProjectsCount  
##              -1.5630                0.3742               -0.1279  
##   MaritalDescMarried  MaritalDescSeparated     MaritalDescSingle  
##              -0.8215               -2.4281               -1.1695  
##   MaritalDescWidowed      EngagementSurvey  
##              -0.3164                0.4190  
## 
## Degrees of Freedom: 226 Total (i.e. Null);  219 Residual
## Null Deviance:       288 
## Residual Deviance: 266.9     AIC: 282.9

Training model

Final Logistic Regression Model is:

model <- glm(formula = Termd ~ DaysLateLast30 + SpecialProjectsCount + 
    MaritalDesc + EngagementSurvey, family = "binomial", data = train)

By using function summary() we obtain the results of our model:

summary(model)

## 
## Call:
## glm(formula = Termd ~ DaysLateLast30 + SpecialProjectsCount + 
##     MaritalDesc + EngagementSurvey, family = "binomial", data = train)
## 
## Deviance Residuals: 
##     Min       1Q   Median       3Q      Max  
## -1.7512  -0.9092  -0.6677   1.1793   2.1308  
## 
## Coefficients:
##                      Estimate Std. Error z value Pr(>|z|)   
## (Intercept)          -1.56299    1.12227  -1.393   0.1637   
## DaysLateLast30        0.37418    0.13826   2.706   0.0068 **
## SpecialProjectsCount -0.12791    0.07243  -1.766   0.0774 . 
## MaritalDescMarried   -0.82153    0.47621  -1.725   0.0845 . 
## MaritalDescSeparated -2.42813    1.13615  -2.137   0.0326 * 
## MaritalDescSingle    -1.16948    0.47768  -2.448   0.0144 * 
## MaritalDescWidowed   -0.31640    1.48437  -0.213   0.8312   
## EngagementSurvey      0.41904    0.24107   1.738   0.0822 . 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 288.04  on 226  degrees of freedom
## Residual deviance: 266.93  on 219  degrees of freedom
## AIC: 282.93
## 
## Number of Fisher Scoring iterations: 4

The AIC or Akaike Information Criteria in this case is 288.74 . The AIC is the measure of fit which penalizes model for the number of model coefficients.

-The p-value in the last column is more than 0.05 for the variables “EngagementSurvey” and “MaritalDescWidowed”, we consider them to be statistically insignificant . As for DaysLateLast30 , MaritalDescSingle and MaritalDescSeparated has the lowest p-value impacts the “Termd” value in this regression model.

-In the case of DaysLateLast30, we see that the estimate is positive, meaning that if employees who been late to work frequently during the last 30 days are significantly more likely to leave.

-Our MaritalDescSeparated and MaritalDescSingle coefficients are significant and negative. This means single and separate employees are less likely to leave than married employees .

Predicting with test Data

testpredict<- predict(model, test, type = "response")

hist(testpredict, main = "Distribution of Test Set \nPredicted Probabilities", 
     xlab = "Probability termination", col = cbPalette[4], border = F, breaks = 15)
abline(v = .5, col = "red", lwd = 3)

prop.table(table(testpredict >= .5))

## 
##     FALSE      TRUE 
## 0.8684211 0.1315789

newPredict <- data.frame(DaysLateLast30 = 0 , SpecialProjectsCount = 5  , MaritalDesc="Married"  , EngagementSurvey=4.96)
testpredict1<- predict(model, newPredict, type = "response")
testpredict1

##         1 
## 0.2797673

#validate the model confusion matrix
prop.table(table(test$Termd))

## 
##         0         1 
## 0.6184211 0.3815789

accuracy <- table(testpredict > .5, test$Termd) # confusion matrix

addmargins(table(testpredict > .5, test$Termd))

##        
##          0  1 Sum
##   FALSE 42 24  66
##   TRUE   5  5  10
##   Sum   47 29  76

(accuracy[[1,1]] + accuracy [[2,2]]) / sum(accuracy)

## [1] 0.6184211

Cutoff

addmargins(accuracy)

##        
##          0  1 Sum
##   FALSE 42 24  66
##   TRUE   5  5  10
##   Sum   47 29  76

With this cutoff, our model catches 5 of the 29 termd (leave) events. This gives us a True Positive Rate of 17% . That is, we are correctly labeling 5 of the 29 “termination” cases. Correspondingly, that same .5 cutoff catches 42 of the 47 stay(not terminated) events, yielding a False Negative Rate of 89% percent.

ROC AUC Curve

‘Receiver Operating Characteristic’ or ROC curve gives an idea on the performance of a model by measuring the trade off between true-positive rate and false-positive rate. Higher the area under the curve(AUC), better is the model.

The model is evaluated using the Confusion matrix, AUC(Area under the curve), and ROC(Receiver operating characteristics) curve.

First I’m going to install some necessary packages for this part.

#install.packages("ROCR")
#install.packages("caTools")

# Loading packages
library(ROCR) # For ROC curve to evaluate model 
library(caTools) # For Logistic regression ,calculation of AUC

pr <- prediction(testpredict, test$Termd)
prf <- performance(pr,measure = "tpr", x.measure = "fpr")

# Function to get the best cutoff point
opt.cut <- function(prf, pr){
    cut.ind <- mapply(FUN=function(x, y, p){
        d <- (x - 0)^2 + (y-1)^2
        ind <- which(d == min(d))
        c(sensitivity = y[[ind]], specificity = 1-x[[ind]], 
            cutoff = p[[ind]])
    }, prf@x.values, prf@y.values, pr@cutoffs)
}

# the plot
plot(prf, colorize = TRUE, print.cutoffs.at = seq(0,1,.1), 
         main = "ROC Curve", lwd = 2)
    abline(coef = c(0,1), col = "black", lwd = 2)
    
# get the optimum cutoff
opt <- opt.cut(prf, pr)
    points(x = 1-opt[2], y = opt[1], pch = 19, col = "red", cex = 1.5)
    text(x = 1-opt[2],  y = opt[1] + .05, labels = "Optimum Cutoff")
    
# Area Under the Curve
text(x = .6, y = .3, label = paste("Area Under the Curve:\n", 
                                       round(as.numeric(performance(pr, "auc")@y.values), 2)))
    
    text(x = .6, y = .15, label = paste("Optimum Cutoff:\n",  round(opt[3],3)))

Interpreting the ROC Curve

We can see that the AUC is 0.6555, which is high. This indicates that our model does a good but not excellent job of predicting whether or not an employee is going to terminate(leave). The more AUC is, the better the model performs. We can see values for the whole continuum of predicted probabilities. The color-coded line tells you what the cutoff values are at each point along the curve. This graph definitely proves our model is good because it curves substantially above the diagonal line.

The best possible cutoff value is 0.324

To get the best trade off between False Positives and True Positives, we would categorize everyone below 0.324 as a “0” (stayer) and everyone at or above it as 1 (leaver).